This paper describes a new approach in the development of training- and testing software for Continuous Hidden Markov Models (CHMM) that has been successfully implemented and utilized by the ESPRIT E project 2094, SUNSTAR. The basic idea behind the new software, called SIRtrain, has been to adopt existing de-facto standards from another ESPRIT project, SAM, thus enabling SIRtrain to work in conjunction with already available speech databases and software tools. Also, the use of a modular and object oriented approach contributes to making SIRtrain a general platform for testing out new principles within speech recognition. Being written in C++ and running on ordinary SUN SPARCstations without additional hardware, the software allows researchers to implement and test new ideas without having to invest in expensive hardware equipment.